M2AttR2U-net: Networks with Multiple Inputs and Multi-layer Supervision for the Bridge Cracks

Numerous attempts have been made to tackle the challenge of defect segmentation, whether through traditional computer vision algorithms or deep learning approaches. However, issues such as discontinuous segmentation lines and the failure to identify small defects still persist. In this paper, we present a novel defect segmentation network that leverages multiple inputs and multiple layers of supervised networks. Specifically, we utilize traditional computer vision algorithms to preprocess image data for input expansion, while simultaneously designing a multi-layer loss function for output supervision. This approach enables end-to-end training and yields favorable results using lower quality images. Notably, our method achieves an mIOU performance of 91.4 on a bridge crack dataset.